Even if the exchange does not advertise the exchange LAN, it's probably
First of all I agree with Leo that not advertising IX prefixes permanently causes more problems than it solves. the case that it is in the IGP (or at least IBGP) of everyone connected to it Well if I would peer with such an ISP at London and Frankfurt I could create a GRE tunnel from London to Frankfurt via the other ISP and use it to transport packets that would otherwise have to traverse my backbone. Or if my peer has a router at IX that happens to have full routing view I can just point a static default toward it and have a free transit. Check out: http://www.bcp38.info adam -----Original Message----- From: Leo Bicknell [mailto:bicknell@ufp.org] Sent: Thursday, April 04, 2013 9:29 PM To: NANOG Subject: Re: route for linx.net in Level3? In a message written on Thu, Apr 04, 2013 at 02:57:11PM -0400, Jay Ashworth wrote:
Yes. In the fallout from the Cloudflare attack of last week it was announced that several IXs were going to stop advertising the address space of their peering lan, which properly does not need to be advertised anyway.
Well, now that's a big maybe. I was a big advocate for the peering exchanges each having their own ASN and announcing the peering block back in the day, and it seems people may have forgotten some of the issues with unadvertised peering exchange blocks. It breaks traceroute for many people: The ICMP TTL Unreachable will come from a non-routed network (the exchange LAN). If it crosses another network boundary doing uRPF, even in loose mode, those unreachables will be dropped. It also reduces the utility of a tool like MTR. Without the ICMP responese it won't know where to ping, and even if it receives the ICMP it's likely packets towards the LAN IP's will be dropped with no route to host. It has the potential to break PMTU discovery for many people: If a router is connected to the exchange and a lower MTU link a packet coming in with DF set will get an ICMP would-fragment reply. Most vendors source from the input interface, e.g. the exchange IP. Like the traceorute case, if crosses another network boundary doing uRPF, even in loose mode, those ICMP messages will be lost, resulting in a PMTU black hole. Some vendors have knobs to force the ICMP to be emitted from a loopback, but not all. People would have to turn it on. But hey, this is a good thing because a DDOS caused issues, right? Well, not so much. Even if the exchange does not advertise the exchange LAN, it's probably the case that it is in the IGP (or at least IBGP) of everyone connected to it, and by extension all of their customers with a default route pointed at them. For the most popular exchanges (AMS-IX, for instance) I suspect the percentage of end users who can reach the exchange LAN without it being explicitly routed to be well over 80%, perhaps into the upper 90% range. So when those boxes DDOS, they are going to all DDOS the LAN anyway. Security through obscurity does not work. This is going to annoy some people just trying to do their day job, and not make a statistical difference to the attackers trying to take out infrastructure. How about we all properly implement BCP 38 instead? -- Leo Bicknell - bicknell@ufp.org - CCIE 3440 PGP keys at http://www.ufp.org/~bicknell/